sched_ext: Fix spurious WARN on stale ops_state in ops_dequeue()

sched_ext: Fix spurious WARN on stale ops_state in ops_dequeue()

ops_dequeue() can race with finish_dispatch() and spuriously trigger the
"queued task must be in BPF scheduler's custody" warning.

ops_dequeue() snapshots p->scx.ops_state via atomic_long_read_acquire()
and then, in the SCX_OPSS_QUEUED arm, asserts that SCX_TASK_IN_CUSTODY
is set. The two reads are not atomic w.r.t. a concurrent
finish_dispatch() running on another CPU:

CPU 1                                    CPU 2
=====                                    =====
                                         dequeue_task_scx()
                                           ops_dequeue()
                                             opss = read_acquire(ops_state)
                                                  = SCX_OPSS_QUEUED
finish_dispatch()
  cmpxchg ops_state:
    SCX_OPSS_QUEUED -> SCX_OPSS_DISPATCHING  [succeeds]
  dispatch_enqueue(SCX_DSQ_GLOBAL,
                   SCX_ENQ_CLEAR_OPSS)
    call_task_dequeue()
      p->scx.flags &= ~SCX_TASK_IN_CUSTODY
                                             WARN_ON_ONCE(!(p->scx.flags &
                                                     SCX_TASK_IN_CUSTODY))
                                            /* opss is stale: QUEUED,
                                             * but task already claimed */
    set_release(ops_state, SCX_OPSS_NONE)

The race has been observed via two distinct call chains: the most common
goes through sched_setaffinity(), a rarer variant through
sched_change_begin().

For SCX_DSQ_GLOBAL / SCX_DSQ_BYPASS, dispatch_enqueue() clears
SCX_TASK_IN_CUSTODY before clearing ops_state to SCX_OPSS_NONE
(intentional, to avoid concurrent non-atomic RMW of p->scx.flags against
ops_dequeue()). The window between those two writes is exactly what
ops_dequeue() observes as "QUEUED without custody".

The observed state is not actually inconsistent, it just means CPU 1 has
already claimed the task and the QUEUED value held by CPU 2 is stale.
Re-read ops_state in that case; the next read is guaranteed to return
SCX_OPSS_DISPATCHING or SCX_OPSS_NONE, both of which exit the switch
cleanly. The retry is bounded: once IN_CUSTODY is cleared, ops_state has
already advanced past QUEUED for this dispatch cycle, and a fresh QUEUED
would require re-enqueue under p's rq lock, which CPU 2 holds.

Changes in v2:
- Use READ_ONCE() for p->scx.flags to ensure fresh reads and prevent
  compiler reordering in the lockless path
- Add cpu_relax() to reduce power consumption and improve performance
  during the spin-wait
- Use unlikely() to optimize branch prediction for the common case
- Expand the in-code comment to document the race condition and
  bounded retry guarantee

Fixes: ebf1ccff79c4 ("sched_ext: Fix ops.dequeue() semantics")
Suggested-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Samuele Mariotti <smariotti@disroot.org>
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Signed-off-by: Tejun Heo <tj@kernel.org>